QuasiNewton is often the faster method (requires fewer steps), but SteepestDescent can converge in situations where QuasiNewton has problems, e.g. if it gets stuck.
The geometric constraints are specified in VNL in the NanoLanguage Scripter, on the "Configuration" tab, and are given just as in NanoLanguage, i.e. you provide the indices of the atoms in the central region that should not move in the relaxation.